- Démarrage
- Agents UiPath dans Studio Web
- À propos des agents UiPath
- Licences
- Exécuter des agents
- Agents et workflows
- Meilleures pratiques pour créer des agents
- Choix du meilleur modèle pour votre agent
- Meilleures pratiques pour la publication et le déploiement d’agents
- Meilleures pratiques en matière d’ingénierie contextuelle
- Meilleures pratiques pour DeepRAG et la transformation par lots : JIT vs. stratégies basées sur l’index
- Utiliser des fichiers
- Contextes
- Configuring agent memory
- Escalades
- Évaluations
- Traçages d’agent
- Score de l’agent
- Gérer les agents UiPath
- Agents codés UiPath
- Créer avec des agents codés
Create a memory space, enable episodic and escalation memory on an agent, and test recall before the agent runs in production.
Agent memory is set up in two places. A memory space is created in the Agents instance management page, then attached to an agent and tuned in Agent Builder in Studio Web. For what each memory type does and when to use it, refer to Agent memory.
Create a memory space
A memory space holds the memory items for one or more agents. Agents that share a space also share what they learn from each other's cases.
- In Agents, go to Management and select Memory spaces.
- Select Create memory space.
- Enter a Name for the memory space.
- (Optional) Enter a Description, up to 1024 characters.
- Select the Folder that the memory space belongs to.
- (Optional) If your organization uses customer managed keys (CMK), enable encryption.
- Sélectionnez Enregistrer.
The memory space appears in the Memory spaces list and is available to attach to an agent. Encryption cannot be changed after the space is created.
Enable episodic memory
Episodic memory adds context from semantically similar past cases to each run. It is configured on the Episodic tab of the Edit & test memory dialog.
| Paramètre | Ce qu’il contrôle |
|---|---|
| Dynamic few-shot learning | Automatically selects the most relevant past interactions as examples to guide the agent's response. |
| Similitude sémantique | How closely a new query must match past ones to be used as an example. Higher values are stricter. |
| Examples to retrieve (K) | How many examples to include in the dynamic few-shot learning. |
| Search mode | Hybrid combines keyword and meaning-based matching for the broadest results. Semantic matches on meaning alone, ignoring exact keywords. |
| Weight | An input argument's share of the match. Weights across all arguments always add up to 100%. |
Give a higher weight to arguments that define the case, such as a customer ID, and a lower weight to arguments that vary freely, such as a free-text note.
- Open your agent in Studio Web.
- Add a Memory node to your agent.
- From the properties panel, select an existing memory space, or create a new one.
- Select Edit / test.
- On the Episodic tab, enable Dynamic few-shot learning.
- Set Semantic similarity, Examples to retrieve (K), and Search mode.
- Under Arguments, select Add argument.
- Choose an input argument from the list.
- Enter a Weight for the argument.
- Repeat steps 7 to 9 for each input that should contribute to the match.
- Sélectionnez Enregistrer.
- Publish the agent.
The configuration applies to runs of the published agent.
Enable escalation memory
Escalation memory reuses how a person resolved a comparable escalation, so the agent applies the stored outcome instead of raising the same escalation again. It is configured on the Escalation tab of the same dialog, and it retrieves only the top-ranked memory.
Escalation memory requires at least one argument under Arguments (step 8).
- Open your agent in Studio Web.
- Add a Memory node to your agent.
- From the properties panel, select an existing memory space, or create a new one.
- Select Edit / test.
- Open the Escalation tab.
- Set Semantic similarity and Search mode.
- Select Add Escalation.
- Under Arguments, select Add argument.
- Choose an input argument from the list.
- Enter a Weight for the argument.
- Sélectionnez Enregistrer.
- Publish the agent.
Test a memory space
The memory playground runs the same recall step the agent would run, without running the agent. Use it to check that your arguments, weights, and similarity threshold return the memories you expect.
- Open your agent in Studio Web.
- In the properties panel, go to the Memory section.
- Select Edit / test.
- Open the Episodic or Escalation tab.
- Enter a test value for each argument.
- Sélectionnez Run (Exécuter).
The playground lists the memories that would be retrieved for that input, along with their similarity scores. On the Escalation tab, the result shows whether a stored resolution would be applied.
Manage memory items
In Agents instance management, open a memory space to view its memory items in a grid. From the grid you can:
- Review each stored item, including the resolution captured for an escalation memory.
- Edit an entry.
- Enable or disable memory items to control whether they are considered during a memory search.
- Delete entries that no longer apply.
- Export entries, and import them into another memory space or tenant.
Editing the inputs of an existing episodic memory item is not supported. To correct one, delete it and add a new one.